Use irtt streaming mode (-s) for unbounded pinger sessions

[ooonea]

Closes #394 — and, per the review discussion, gawk is now gone entirely: one irtt client -s -r --loose -i <interval> per reflector, parsed in the main loop like every other pinger method.

The raw stream carries neither the reflector nor a wall-clock timestamp, all pingers share the one response fd, and irtt client takes exactly one host per session (irtt_client.go:284) with no flag to tag its output — so a per-reflector plain-bash helper prefixes those two fields onto each line and handles the respawn (a server-side max duration clamps the negotiated session and the client just exits; rate-limited to one relaunch per second on quick deaths).

Per-line gating, now in the extraction arm: numeric rd/sd via glob class checks (no [[ =~ ]] on the hot path). That one check drops NaN OWDs (no server timestamps), duplicates (dup=1 lines carry NaN OWDs by construction) and negative OWDs (client and server clocks out of step — which to_us() used to reject, and the interim transcoder let through). ipdv is legitimately NaN on the first reply of every session and on seqno wrap (heistp confirmed the wrap behavior in heistp/irtt#45), so it never gates a sample. ms→µs uses the same printf %.3f idiom as the fping arm.

--loose: without it, any server-imposed restriction (e.g. a public server's min send interval) kills the client pre-test at every respawn — [ServerRestriction] server increased interval from 100ms to 1s, exit 1, zero samples, silent spin. Reproduced both ways against irtt server -i 1s.

Removed: the gawk + time-extension dependencies, the -d session bound, the WaitForPackets match, the unit-suffix parsing, and stdbuf (raw lines come from Go's unbuffered stdout — the old stdbuf -oL never had an effect for the same reason). In raw mode irtt routes events/notices to stderr, so stdout carries only sample lines. ping_prefix_string now wraps the irtt client itself, as in the other arms (it used to wrap gawk; the popen'd irtt never carried the prefix).

Validation: the gawk checks are gone; a time-bounded probe of irtt client -h checks for streaming-mode support — -s only exists since irtt 0.9.2, and without the probe an older irtt would reject -s at every respawn while the daemon silently produces no samples (min-rate pinning). No pipe to grep (the help goes to stderr and exits 2, which pipefail would turn into a false failure); read -t bounds it so a never-exiting irtt can't hang startup.

irtt_session_duration_m: deprecated-and-ignored rather than dropped, since dropping the key would break existing config files that set it — happy to drop outright instead, your call from the issue.

Also worth noting from heistp/irtt#45: a 1-hour 1 ms-interval stream (3.6M round trips) shows no leak, RSS stable ~12.5 MB.

Tested on a loopback rig (real irtt 0.9.2 server, wrapper and parser blocks extracted verbatim from the file): 9-field lines end-to-end, µs values matching irtt's own rd/sd within rounding, first sample kept, dup/NaN/negative/short/long lines dropped, respawn paced ~1/s across server restarts, group kill leaves no orphans, stdout stays pure across a server bounce, capability gate passes on 0.9.2 / fails on older / fails-not-hangs on a stuck binary; shellcheck finding count identical to master. Hot-path cost measured over 1M iterations: extraction arm 5.8→20.8 µs/sample (the ms→µs conversion moved in — same class of work the fping arm already does), helper 9.6 µs/line in its own background process vs 4.2 for the gawk stage; at -i 0.3s × 6 pingers that is under 0.1% of a core either way. Net diff vs master: −5 lines.

[lynxthecat]

But since irtt now outputs a stream like ping, tsping and fping, can't we forego gawk altogether and do what we do with ping, tsping and fping?

Why can't we do something much simpler like:

		fping)
			${ping_prefix_string} fping ${ping_extra_args} --timestamp --loop --period "${reflector_ping_interval_ms}" --interval "${ping_response_interval_ms}" --timeout 10000 "${reflectors[@]:0:${no_pingers}}" 2> /dev/null >&"${main_fd}" &
			pinger_pids[0]=${!}
			proc_pids['fping_pinger']=${pinger_pids[0]}

I'd like to dispense with gawk assuming it is no longer necessary and bring irtt in line with the ping, fping and tsping implementations.

[ooonea]

Done — gawk is gone entirely; the extraction arm now parses the raw stream like ping/tsping/fping (glob class checks, ms→µs via the same printf %.3f idiom as the fping arm). PR body updated to match.

One wrinkle kept me from dropping the helper process too: the raw lines carry neither the reflector nor a timestamp, all pingers share the one response fd, and irtt client takes exactly one host per session (irtt_client.go:284) with no flag to tag its output — so a per-reflector plain-bash loop still prefixes those two fields and handles the respawn. Plain bash, no gawk, no time extension.

Testing against a real 0.9.2 server surfaced two fixes now included: --loose (without it a server restriction, e.g. a public server's min send interval, kills the client pre-test at every respawn — silent zero-sample spin), and the glob gate rejecting negative rd/sd (unsynced clocks), which to_us() used to drop but the interim transcoder let through.

Hot-path numbers are in the PR body; short version: under 0.1% of a core either way at -i 0.3s × 6 pingers, and the diff vs master is now −5 lines with no non-ping dependency left on the irtt path.